Given this trend, GPT 5 or 6 will be trained on a majority of content from its previous versions, modeling them instead of the expect full range of language. Researchers have already tested the outcome of a model-in-loop with pictures, it was not pretty.
A process owned by any user will be able to exploit a userspace vulnerability, whatever this user is. Selinux, chroot, cgroups/containerization add a layer of protection to this, but any vulnerability that bypass these will be as exploitable from nobody as from any other local user. It will protect a user files from some access attempts but will fail to prevent any serious attack. And as usual when it comes to security, a false sense of security is worse than no security at all.
Remember that some exploits exist that can climb outside of a full-blown virtual machine to the virtualisation host, finding a user escalation vulnerability is even more likely.
The only real protection is an up-to-date system, sane user behavior and maybe a little bit of paranoia.